A Survey of Outlier Detection Methodologies.

Size: px
Start display at page:

Download "A Survey of Outlier Detection Methodologies."


1 A Survey of Outlier Detection Methodologies. Victoria J. Hodge Dept. of Computer Science, University of York, York, YO10 5DD UK tel: fax: and Jim Austin Abstract. Outlier detection has been used for centuries to detect and, where appropriate, remove anomalous observations from data. Outliers arise due to mechanical faults, changes in system behaviour, fraudulent behaviour, human error, instrument error or simply through natural deviations in populations. Their detection can identify system faults and fraud before they escalate with potentially catastrophic consequences. It can identify errors and remove their contaminating effect on the data set and as such to purify the data for processing. The original outlier detection methods were arbitrary but now, principled and systematic techniques are used, drawn from the full gamut of Computer Science and Statistics. In this paper, we introduce a survey of contemporary techniques for outlier detection. We identify their respective motivations and distinguish their advantages and disadvantages in a comparative review. Keywords: Outlier, Novelty, Anomaly, Noise, Deviation, Detection, Recognition 1. Introduction Outlier detection encompasses aspects of a broad spectrum of techniques. Many techniques employed for detecting outliers are fundamentally identical but with different names chosen by the authors. For example, authors describe their various approaches as outlier detection, novelty detection, anomaly detection, noise detection, deviation detection or exception mining. In this paper, we have chosen to call the technique outlier detection although we also use novelty detection where we feel appropriate but we incorporate approaches from all five categories named above. Additionally, authors have proposed many definitions for an outlier with seemingly no universally accepted definition. We will take the definition of Grubbs (Grubbs, 1969) and quoted in Barnett & Lewis (Barnett and Lewis, 1994): An outlying observation, or outlier, is one that appears to deviate markedly from other members of the sample in which it occurs. This work was supported by EPSRC Grant No. GR/R55191/01 c 2004 Kluwer Academic Publishers. Printed in the Netherlands. Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.1

2 2 A further outlier definition from Barnett & Lewis (Barnett and Lewis, 1994) is: An observation (or subset of observations) which appears to be inconsistent with the remainder of that set of data. In figure 2, there are five outlier points labelled V, W, X, Y and Z which are clearly isolated and inconsistent with the main cluster of points. The data in the figures in this survey paper is adapted from the Wine data set (Blake+Merz, 1998). John (John, 1995) states that an outlier may also be surprising veridical data, a point belonging to class A but actually situated inside class B so the true (veridical) classification of the point is surprising to the observer. Aggarwal (Aggarwal and Yu, 2001) notes that outliers may be considered as noise points lying outside a set of defined clusters or alternatively outliers may be defined as the points that lie outside of the set of clusters but are also separated from the noise. These outliers behave differently from the norm. In this paper, we focus on the two definitions quoted from (Barnett and Lewis, 1994) above and do not consider the dual class-membership problem or separating noise and outliers. Outlier detection is a critical task in many safety critical environments as the outlier indicates abnormal running conditions from which significant performance degradation may well result, such as an aircraft engine rotation defect or a flow problem in a pipeline. An outlier can denote an anomalous object in an image such as a land mine. An outlier may pinpoint an intruder inside a system with malicious intentions so rapid detection is essential. Outlier detection can detect a fault on a factory production line by constantly monitoring specific features of the products and comparing the real-time data with either the features of normal products or those for faults. It is imperative in tasks such as credit card usage monitoring or mobile phone monitoring to detect a sudden change in the usage pattern which may indicate fraudulent usage such as stolen card or stolen phone airtime. Outlier detection accomplishes this by analysing and comparing the time series of usage statistics. For application processing, such as loan application processing or social security benefit payments, an outlier detection system can detect any anomalies in the application before approval or payment. Outlier detection can additionally monitor the circumstances of a benefit claimant over time to ensure the payment has not slipped into fraud. Equity or commodity traders can use outlier detection methods to monitor individual shares or markets and detect novel trends which may indicate buying or selling opportunities. A news delivery system Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.2

3 can detect changing news stories and ensure the supplier is first with the breaking news. In a database, outliers may indicate fraudulent cases or they may just denote an error by the entry clerk or a misinterpretation of a missing value code, either way detection of the anomaly is vital for data base consistency and integrity. A more exhaustive list of applications that utilise outlier detection is: 3 Fraud detection - detecting fraudulent applications for credit cards, state benefits or detecting fraudulent usage of credit cards or mobile phones. Loan application processing - to detect fraudulent applications or potentially problematical customers. Intrusion detection - detecting unauthorised access in computer networks. Activity monitoring - detecting mobile phone fraud by monitoring phone activity or suspicious trades in the equity markets. Network performance - monitoring the performance of computer networks, for example to detect network bottlenecks. Fault diagnosis - monitoring processes to detect faults in motors, generators, pipelines or space instruments on space shuttles for example. Structural defect detection - monitoring manufacturing lines to detect faulty production runs for example cracked beams. Satellite image analysis - identifying novel features or misclassified features. Detecting novelties in images - for robot neotaxis or surveillance systems. Motion segmentation - detecting image features moving independently of the background. Time-series monitoring - monitoring safety critical applications such as drilling or high-speed milling. Medical condition monitoring - such as heart-rate monitors. Pharmaceutical research - identifying novel molecular structures. Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.3

4 4 Detecting novelty in text - to detect the onset of news stories, for topic detection and tracking or for traders to pinpoint equity, commodities, FX trading stories, outperforming or under performing commodities. Detecting unexpected entries in databases - for data mining to detect errors, frauds or valid but unexpected entries. Detecting mislabelled data in a training data set. Outliers arise because of human error, instrument error, natural deviations in populations, fraudulent behaviour, changes in behaviour of systems or faults in systems. How the outlier detection system deals with the outlier depends on the application area. If the outlier indicates a typographical error by an entry clerk then the entry clerk can be notified and simply correct the error so the outlier will be restored to a normal record. An outlier resulting from an instrument reading error can simply be expunged. A survey of human population features may include anomalies such as a handful of very tall people. Here the anomaly is purely natural, although the reading may be worth flagging for verification to ensure no errors, it should be included in the classification once it is verified. A system should use a classification algorithm that is robust to outliers to model data with naturally occurring outlier points. An outlier in a safety critical environment, a fraud detection system, an image analysis system or an intrusion monitoring system must be detected immediately (in real-time) and a suitable alarm sounded to alert the system administrator to the problem. Once the situation has been handled, this anomalous reading may be stored separately for comparison with any new fraud cases but would probably not be stored with the main system data as these techniques tend to model normality and use this to detect anomalies. There are three fundamental approaches to the problem of outlier detection: 1. Type 1 - Determine the outliers with no prior knowledge of the data. This is essentially a learning approach analogous to unsupervised clustering. The approach processes the data as a static distribution, pinpoints the most remote points, and flags them as potential outliers. Type 1 assumes that errors or faults are separated from the normal data and will thus appear as outliers. In figure3,pointsv,w,x,yandzaretheremotepointsseparated from the main cluster and would be flagged as possible outliers. We note that the main cluster may be subdivided if necessary into more Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.4

5 than one cluster to allow both classification and outlier detection as with figure 4. The approach is predominantly retrospective and is analogous to a batch-processing system. It requires that all data be available before processing and that the data is static. However, once the system possesses a sufficiently large database with good coverage, then it can compare new items with the existing data. There are two sub-techniques commonly employed, diagnosis and accommodation (Rousseeuw and Leroy, 1996). An outlier diagnostic approach highlights the potential outlying points. Once detected, the system may remove these outlier points from future processing of the data distribution. Many diagnostic approaches iteratively prune the outliers and fit their system model to the remaining data until no more outliers are detected. An alternative methodology is accommodation that incorporates the outliers into the distribution model generated and employs a robust classification method. These robust approaches can withstand outliers in the data and generally induce a boundary of normality around the majority of the data which thus represents normal behaviour. In contrast, non-robust classifier methods produce representations which are skewed when outliers are left in. Non-robust methods are best suited when there are only a few outliers in the data set (as in figure 3) as they are computationally cheaper than the robust methods but a robust method must be used if there are a large number of outliers to prevent this distortion. Torr & Murray (Torr and Murray, 1993) use a cheap Least Squares algorithm if there are only a few outliers but switch to a more expensive but robust algorithm for higher frequencies of outliers. 2. Type 2 - Model both normality and abnormality. This approach is analogous to supervised classification and requires pre-labelled data, tagged as normal or abnormal. In figure 4, there are three classes of normal data with pre-labelled outliers in isolated areas. The entire area outside the normal class represents the outlier class. The normal points could be classified as a single class or subdivided into the three distinct classes according to the requirements of the system to provide a simple normal/abnormal classification or to provide an abnormal and 3-classes of normally classifier. Classifiers are best suited to static data as the classification needs to be rebuilt from first principles if the data distribution shifts unless the system uses an incremental classifier such as an evolutionary neural network. We describe one such approach later which 5 Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.5

6 6 uses a Grow When Required (Marsland, 2001) network. A type 2 approach can be used for on-line classification, where the classifier learns the classification model and then classifies new exemplars as and when required against the learned model. If the new exemplar lies in a region of normality it is classified as normal, otherwise it is flagged as an outlier. Classification algorithms require a good spread of both normal and abnormal data, i.e., the data should cover the entire distribution to allow generalisation by the classifier. New exemplars may then be classified correctly as classification is limited to a known distribution and a new exemplar derived from a previously unseen region of the distribution may not be classified correctly unless the generalisation capabilities of the underlying classification algorithm are good. 3. Type 3 - Model only normality or in a very few cases model abnormality (Fawcett and Provost, 1999), (Japkowicz et al., 1995). Authors generally name this technique novelty detection or novelty recognition. It is analogous to a semi-supervised recognition or detection task and can be considered semi-supervised as the normal class is taught but the algorithm learns to recognise abnormality. The approach needs pre-classified data but only learns data marked normal. It is suitable for static or dynamic data as it only learns one class which provides the model of normality. It can learn the model incrementally as new data arrives, tuning the model to improve the fit as each new exemplar becomes available. It aims to define a boundary of normality. A type 3 system recognises a new exemplar as normal if it lies within the boundary and recognises the new exemplar as novel otherwise. In figure 5, the novelty recogniser has learned the same data as shown in figure 2 but only the normal class is learned and a boundary of normality induced. If points V, W, X, Y and Z from figure 2 are compared to the novelty recogniser they will be labelled as abnormal as they lie outside the induced boundary. This boundary may be hard where a point lies wholly within or wholly outside the boundary or soft where the boundary is graduated depending on the underlying detection algorithm. A soft bounded algorithm can estimate the degree of outlierness. It requires the full gamut of normality to be available for training to permit generalisation. However, it requires no abnormal data for training unlike type 2. Abnormal data is often difficult to obtain or expensive in many fault detection domains such as aircraft engine Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.6

7 monitoring. It would be extremely costly to sabotage an aircraft engine just to obtain some abnormal running data. Another problem with type 2 is it cannot always handle outliers from unexpected regions, for example, in fraud detection a new method of fraud never previously encountered or previously unseen fault in a machine may not be handled correctly by the classifier unless generalisation is very good. In this method, as long as the new fraud lies outside the boundary of normality then the system will be correctly detect the fraud. If normality shifts then the normal class modelled by the system may be shifted by re-learning the data model or shifting the model if the underlying modelling technique permits such as evolutionary neural networks. The outlier approaches described in this survey paper generally map data onto vectors. The vectors comprise numeric and symbolic attributes to represent continuous-valued, discrete (ordinal), categorical (unordered numeric), ordered symbolic or unordered symbolic data. The vectors may be monotype or multi-type. The statistical and neural network approaches typically require numeric monotype attributes and need to map symbolic data onto suitable numeric values 1 but the machine learning techniques described are able to accommodate multitype vectors and symbolic attributes. The outliers are determined from the closeness of vectors using some suitable distance metric. Different approaches work better for different types of data, for different numbers of vectors, for different numbers of attributes, according to the speed required and according to the accuracy required. The two fundamental considerations when selecting an appropriate methodology for an outlier detection system are: 7 selecting an algorithm which can accurately model the data distribution and accurately highlight outlying points for a clustering, classification or recognition type technique. The algorithm should also be scalable to the data sets to be processed. selecting a suitable neighbourhood of interest for an outlier. The selection of the neighbourhood of interest is non-trivial. Many algorithms define boundaries around normality during processing and autonomously induce a threshold. However, these approaches are often parametric enforcing a specific distribution model or require user-specified parameters such as the number of clusters. Other techniques discussed below require user-defined parameters to define the size or density of neighbourhoods for outlier 1 There are statistical models for symbolic data (such as log-linear models) but these are not generally used for outlier analysis Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.7

8 8 thresholding. The choice of neighbourhood whether user-defined or autonomously induced needs to be applicable for all density distributions likely to be encountered and can potentially include those with sharp density variations. In the remainder of this paper, we categorise and analyse broad range of outlier detection methodologies. We pinpoint how each handles outliers and make recommendations for when each methodology is appropriate for clustering, classification and/or recognition. Barnett & Lewis (Barnett and Lewis, 1994) and Rousseeuw & Leroy (Rousseeuw and Leroy, 1996) describe and analyse a broad range of statistical outlier techniques and Marsland (Marsland, 2001) analyses a wide range of neural methods. We have observed that outlier detection methods are derived from three fields of computing: statistics (proximity-based, parametric, non-parametric and semi-parametric), neural networks (supervised and unsupervised) and machine learning. In the next 4 sections, we describe and analyse techniques from all three fields and a collection of hybrid techniques that utilise algorithms from multiple fields. The approaches described here encompass distance-based, set-based, density-based, depth-based, model-based and graph-based algorithms. 2. Statistical models Statistical approaches were the earliest algorithms used for outlier detection. Some of the earliest are applicable only for single dimensional data sets. In fact, many of the techniques described in both (Barnett and Lewis, 1994) and (Rousseeuw and Leroy, 1996) are single dimensional or at best univariate. One such single dimensional method is Grubbs method (Extreme Studentized Deviate) (Grubbs, 1969) which calculates a Z value as the difference between the mean value for the attribute and the query value divided by the standard deviation for the attribute where the mean and standard deviation are calculated from all attribute values including the query value. The Z value for the query is compared with a 1% or 5% significance level. The technique requires no user parameters as all parameters are derived directly from data. However, the technique is susceptible to the number of exemplars in the data set. The higher the number of records the more statistically representative the sample is likely to be. Statistical models are generally suited to quantitative real-valued data sets or at the very least quantitative ordinal data distributions where the ordinal data can be transformed to suitable numerical values for statistical (numerical) processing. This limits their applicability and Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.8

9 increases the processing time if complex data transformations are necessary before processing. Probably one of the simplest statistical outlier detection techniques described here, Laurikkala et al. (Laurikkala et al., 2000) use informal box plots to pinpoint outliers in both univariate and multivariate data sets. This produces a graphical representation (see figure 1 for an example box plot) and allows a human auditor to visually pinpoint the outlying points. It is analogous to a visual inspection of figure 2. Their approach can handle real-valued, ordinal and categorical (no order) attributes. Box plots plot the lower extreme, lower quartile, median, upper quartile and upper extreme points. For univariate data, this is a simple 5-point plot, as in figure 1. The outliers are the points beyond the lower and upper extreme values of the box plot, such as V, Y and Z in figure 1. Laurikkala et al. suggest a heuristic of 1.5 inter-quartile range beyond the upper and lower extremes for outliers but this would need to vary across different data sets. For multivariate data sets the authors note that there are no unambiguous total orderings but recommend using the reduced sub-ordering based on the generalised distance metric using the Mahalanobis distance measure (see equation 2). The Mahalanobis distance measure includes the inter-attribute dependencies so the system can compare attribute combinations. The authors found the approach most accurate for multivariate data where a panel of experts agreed with the outliers detected by the system. For univariate data, outliers are more subjective and may be naturally occurring, for example the heights of adult humans, so there was generally more disagreement. Box plots make no assumptions about the data distribution model but are reliant on a human to note the extreme points plotted on the box plot. Statistical models use different approaches to overcome the problem of increasing dimensionality which both increases the processing time and distorts the data distribution by spreading the convex hull. Some methods preselect key exemplars to reduce the processing time (Datta and Kibler, 1995), (Skalak, 1994). As the dimensionality increases, the data points are spread through a larger volume and become less dense. This makes the convex hull harder to discern and is known as the Curse of Dimensionality. The most efficient statistical techniques automatically focus on the salient attributes and are able to process the higher number of dimensions in tractable time. However, many techniques such as k-nn, neural networks, Minimum Volume Ellipsoid or Convex Peeling described in this survey are susceptible to the Curse of Dimensionality. These approaches may utilise a preprocessing al- 9 Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.9

10 10 gorithm to preselect the salient attributes (Aha and Bankert, 1994), (Skalak, 1994), (Skalak and Rissland, 1990). These feature selection techniques essentially remove noise from the data distribution and focus the main cluster of normal data points while isolating the outliers as with figure 2. Only a few attributes usually contribute to the deviation of an outlier case from a normal case. An alternative technique is to use an algorithm to project the data onto a lower dimensional subspace to compact the convex hull (Aggarwal and Yu, 2001) or use Principal Component Analysis (Faloutsos et al., 1997), (Parra et al., 1996) Proximity-based Techniques Proximity-based techniques are simple to implement and make no prior assumptions about the data distribution model. They are suitable for both type 1 and type 2 outlier detection. However, they suffer exponential computational growth as they are founded on the calculation of the distances between all records. The computational complexity is directly proportional to both the dimensionality of the data m and the number of records n. Hence, methods such as k-nearest neighbour (also known as instance-based learning and described next) with O(n 2 m) runtime are not feasible for high dimensionality data sets unless the running time can be improved. There are various flavours of k-nearest Neighbour (k-nn) algorithm for outlier detection but all calculate the nearest neighbours of a record using a suitable distance calculation metric such as Euclidean distance or Mahalanobis distance. Euclidean distance is given by equation 1 n (x i y i ) 2 (1) i=1 and is simply the vector distance whereas the Mahalanobis distance given by equation 2 (x µ) T C 1 (x µ) (2) calculates the distance from a point to the centroid (µ) defined by correlated attributes given by the Covariance matrix (C). Mahalanobis distance is computationally expensive to calculate for large high dimensional data sets compared to the Euclidean distance as it requires a pass through the entire data set to identify the attribute correlations. Ramaswamy et al. (Ramaswamy et al., 2000) introduce an optimised k- NN to produce a ranked list of potential outliers. A point p is an outlier if no more than n 1 other points in the data set have a higher D m Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.10

11 (distance to m th neighbour) where m is a user-specified parameter. In figure 3, V is most isolated followed by X, W, Y then Z so the outlier rank would be V, X, W, Y, Z. This approach is susceptible to the computational growth as the entire distance matrix must be calculated for all points (ALL k-nn ) so Ramaswamy et al. include techniques for speeding the k-nn algorithm such as partitioning the data into cells. If any cell and its directly adjacent neighbours contains more than k points, then the points in the cell are deemed to lie in a dense area of the distribution so the points contained are unlikely to be outliers. If the number of points lying in cells more than a pre-specified distance apart is less than k then all points in the cell are labelled as outliers. Hence, only a small number of cells not previously labelled need to be processed and only a relatively small number of distances need to be calculated for outlier detection. Authors have also improved the running speed of k-nn by creating an efficient index using a computationally efficient indexing structure (Ester and Xu, 1996) with linear running time. Knorr & Ng (Knorr and Ng, 1998) introduce an efficient type 1 k- NN approach. If m of the k nearest neighbours (where m < k) lie within a specific distance threshold d then the exemplar is deemed to lie in a sufficiently dense region of the data distribution to be classified as normal. However, if there are less than m neighbours inside the distance threshold then the exemplar is an outlier. A very similar type 1 approach for identifying land mines from satellite ground images (Byers and Raftery, 1998) is to take the m-th neighbour and find the distance D m. If this distance is less than a threshold d then the exemplar lies in a sufficiently dense region of the data distribution and is classified as normal. However, if the distance is more than the threshold value then the exemplar must lie in a locally sparse area and is an outlier. This has reduced the number of data-specific parameters from Knorr & Ng s (Knorr and Ng, 1998) approach by one as we now have d and m but no k value.infigure3,pointsv,w,x,yandzarerelativelydistant from their neighbours and will have fewer than m neighbours within d and a high D m so both Knorr & Ng and Byers & Raftery classify them as outliers. This is less susceptible to the computational growth than the ALL k-nn approach as only the k nearest neighbours need to be calculated for a new exemplar rather than the entire distance matrix for all points. A type 2 classification k-nn method such as the majority voting approach (Wettschereck, 1994) requires a labelled data set with both normal and abnormal vectors classified. The k-nearest neighbours for the new exemplar are calculated and it is classified according to the 11 Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.11

12 12 majority classification of the nearest neighbours (Wettschereck, 1994). An extension incorporates the distance where the voting power of each nearest neighbour is attenuated according to its distance from the new item (Wettschereck, 1994) with the voting power systematically decreasing as the distance increases. Tang et al. (Tang et al., 2002) introduce an type 1 outlier diagnostic which unifies weighted k-nn with a connectivity-based approach and calculates a weighted distance score rather than a weighted classification. It calculates the average chaining distance (path length) between a point p and its k neighbours. The early distances are assigned higher weights so if a point lies in a sparse region as points V, W, X, Y and Z in figure 4 its nearest neighbours in the path will be relatively distant and the average chaining distance will be high. In contrast, Wettschereck requires a data set with good coverage for both normal and abnormal points. The distribution in figure 4 would cause problems for voted k-nn as there are relatively few examples of outliers and their nearest neighbours, although distant, will in fact be normal points so they are classified as normal. Tang s underlying principle is to assimilate both density and isolation. A point can lie in a relatively sparse region of a distribution without being an outlier but a point in isolation is an outlier. However, the technique is computationally complex with a similar expected run-time to a full k-nn matrix calculation as it relies on calculating paths between all points and their k neighbours. Wettschereck s approach is less susceptible as only the k nearest neighbours are calculated relative to the single new item. Another technique for optimising k-nn is reducing the number of features. These feature subset selectors are applicable for most systems described in this survey, not just the k-nn techniques. Aha & Bankert (Aha and Bankert, 1994) describe a forward sequential selection feature selector which iteratively evaluates feature subsets adding one extra dimension per iteration until the extra dimension entails no improvement in performance. The feature selector demonstrates high accuracy when coupled with an instance-based classifier (nearest-neighbour) but is computationally expensive due to the combinatorial problem of subset selection. Aggarwal & Yu (Aggarwal and Yu, 2001) employ lower dimensional projections of the data set and focus on key attributes. The method assumes that outliers are abnormally sparse in certain lower dimensional projections where the combination of attributes in the projection correlates to the attributes that are deviant. Aggarwal & Yu use an evolutionary search algorithm to determine the projections which has a faster running time than the conventional approach employed by Aha & Bankert. Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.12

13 13 Proximity based methods are also computationally susceptible to the number of instances in the data set as they necessitate the calculation of all vector distances. Datta & Kibler (Datta and Kibler, 1995) use a diagnostic prototype selection to reduce the storage requirements to a few seminal vectors. Skalak (Skalak, 1994) employs both feature selection and prototype selection where a large data set can be stored as a few lower dimensional prototype vectors. Noise and outliers will not be stored as prototypes so the method is robust. Limiting the number of prototypes prevents over-fitting just as pruning a decision tree can prevent over-fitting by limiting the number of leaves (prototypes) stored. However, prototyping must be applied carefully and selectively as it will increase the sparsity of the distribution and the density of the nearest neighbours. A majority voting k-nn technique such as (Wettschereck, 1994) will be less affected but an approach relying on the distances to the nth neighbour (Byers and Raftery, 1998) or counting the number of neighbours within specific distances (Knorr and Ng, 1998) will be strongly affected. Prototyping is in many ways similar to the k-means and k-medoids described next but with prototyping a new instance is compared to the prototypes using conventional k-nearest neighbour whereas k-means and k-medoids prototypes have a kernel with a locally defined radius and the new instance is compared with the kernel boundaries. A prototype approach is also applicable for reducing the data set for neural networks and decision trees. This is particularly germane for nodebased neural networks which require multiple passes through the data set to train so a reduced data set will entail less training steps. Dragon Research (Allan et al., 1998) and Nairac (Nairac et al., 1999) (discussed in section 5) use k-means for novelty detection. Dragon perform on-line event detection to identify news stories concerning new events. Each of the k clusters provides a local model of the data. The algorithm represents each of the k clusters by a prototype vector with attribute values equivalent to the mean value across all points in that cluster. In figure 4, if k is 3 then the algorithm would effectively circumscribe each class (1, 2 and 3) with a hyper-sphere and these 3 hyper-spheres effectively represent and classify normality. K-means requires the user to specify the value of k in advance. Determining the optimum number of clusters is a hard problem and necessitates running the k-means algorithm a number of times with different k values and selecting the best results for the particular data set. However, the k value is usually small compared with the number of records in the Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.13

14 14 data set. Thus, the computational complexity for classification of new instances and the storage overhead are vastly reduced as new vectors need only be compared to k prototype vectors and only k prototype vectors need be stored unlike k-nn where all vectors need be stored and compared to each other. K-means initially chooses random cluster prototypes according to a user-defined selection process, the input data is applied iteratively and the algorithm identifies the best matching cluster and updates the cluster centre to reflect the new exemplar and minimise the sum-of-squares clustering function given by equation 3 K j=1 n S j x n µ j 2 (3) where µ is the mean of the points (x n )inclusters j. Dragon use an adapted similarity metric that incorporates the word count from the news story, the distance to the clusters and the effect the insertion has on the prototype vector for the cluster. After training, each cluster has a radius which is the distance between the prototype and the most distant point lying in the cluster. This radius defines the bounds of normality and is local to each cluster rather than the global distance settings used in many approaches such as (Knorr and Ng, 1998), (Ramaswamy et al., 2000) and (Byers and Raftery, 1998) k-nn approaches. A new exemplar is compared with the k-cluster model. If the point lies outside all clusters then it is an outlier. A very similar partitional algorithm is the k-medoids algorithm or PAM (Partition Around Medoids) which represents each cluster using an actual point and a radius rather than a prototype (average) point and a radius. Bolton & Hand (Bolton and Hand, 2001) use a k-medoids type approach they call Peer Group Analysis for fraud detection. K- medoids is robust to outliers as it does not use optimisation to solve the vector placement problem but rather uses actual data points to represent cluster centres. K-medoids is less susceptible to local minima than standard k-means during training where k-means often converges to poor quality clusters. It is also data-order independent unlike standard k-means where the order of the input data affects the positioning of the cluster centres and Bradley (Bradley et al., 1999) shows that k-medoids provides better class separation than k-means and hence is better suited to a novelty recognition task due to the improved separation capabilities. However, k-means outperforms k-medoids and can handle larger data sets more efficiently as k-medoids can require O(n 2 ) running time Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.14

15 per iteration whereas k-means is O(n). Both approaches can generalise from a relatively small data set. Conversely, the classification accuracy of k-nn, Least Squares Regression or Grubbs method is susceptible to the number of exemplars in the data set like the kernel-based Parzen windows and the node-based supervised neural networks described later as they all model and analyse the density of the input distribution. The data mining partitional algorithm CLARANS (Ng and Han, 1994), is an optimised derivative of the k-medoids algorithm and can handle outlier detection which is achieved as a by-product of the clustering process. It applies a random but bounded heuristic search to find an optimal clustering by only searching a random selection of cluster updates. It requires two user-specified parameters, the value of k and the number of cluster updates to randomly select. Rather than searching the entire data set for the optimal medoid it tests a pre-specified number of potential medoids and selects the first medoid it tests which improves the cluster quality. However, it still has O(n 2 k) running time so is only really applicable for small to medium size data sets. Another proximity-based variant is the graph connectivity method. Shekhar et al. (Shekhar et al., 2001) introduce an approach to traffic monitoring which examines the neighbourhoods of points but from a topologically connected rather than distance-based perspective. Shekhar detects traffic monitoring stations producing sensor readings which are inconsistent with stations in the immediately connected neighbourhood. A station is an outlier if the difference between its sensor value and the average sensor value of its topological neighbours differs by more than a threshold percentage from the mean difference between all nodes and their topological neighbours. This is analogous to calculating the average distance between each point i and its k neighbours avgdi k and then finding any points whose average distance avgd k differs by more than a specified percentage. The technique only considers topologically connected neighbours so there is no prerequisite for specifying k and the number can vary locally depending on the number of connections. However, it is best suited to domains where a connected graph of nodes is available such as analysing traffic flow networks where the individual monitoring stations represent nodes in the connected network Parametric Methods Many of the methods we have just described do not scale well unless modifications and optimisations are made to the standard algorithm. 15 Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.15

16 16 Parametric methods allow the model to be evaluated very rapidly for new instances and are suitable for large data sets; the model grows only with model complexity not data size. However, they limit their applicability by enforcing a pre-selected distribution model to fit the data. If the user knows their data fits such a distribution model then these approaches are highly accurate but many data sets do not fit one particular model. One such approach is Minimum Volume Ellipsoid estimation (MVE) (Rousseeuw and Leroy, 1996) which fits the smallest permissible ellipsoid volume around the majority of the data distribution model (generally covering 50% of the data points). This represents the densely populated normal region shown in figure 2 (with outliers shown) and figure 5 (with outliers removed). A similar approach, Convex Peeling peels away the records on the boundaries of the data distribution s convex hull (Rousseeuw and Leroy, 1996) and thus peels away the outliers. In contrast MVE maintains all points and defines a boundary around the majority of points. In convex peeling, each point is assigned a depth. The outliers will have the lowest depth thus placing them on the boundary of the convex hull and are shed from the distribution model. For example in figure 2, V, W, X, Y and Z would each be assigned the lowest depth and shed during the first iteration. Peeling repeats the convex hull generation and peeling process on the remaining records until it has removed a pre-specified number of records. The technique is a type 1, unsupervised clustering outlier detector. Unfortunately, it is susceptible to peeling away p + 1 points from the data distribution on each iteration and eroding the stock of normal points too rapidly (Rousseeuw and Leroy, 1996). Both MVE and Convex Peeling are robust classifiers that fit boundaries around specific percentages of the data irrespective of the sparseness of the outlying regions and hence outlying data points do not skew the boundary. Both however, rely on a good spread of the data. Figure 2 has few outliers so an ellipsoid circumscribing 50% of the data would omit many normal points from the boundary of normality. Both MVE and convex peeling are only applicable for lower dimensional data sets (Barnett and Lewis, 1994) (usually three dimensional or less for convex peeling) as they suffer the Curse of Dimensionality where the convex hull is stretched as more dimensions are added and the surface becomes too difficult to discern. Torr and Murray (Torr and Murray, 1993) also peel away outlying Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.16

17 points by iteratively pruning and re-fitting. They measure the effect of deleting points on the placement of the Least Squares standard regression line for a diagnostic outlier detector. The LS line is placed to minimise equation 4 n (y i ŷ i ) 2 (4) i=1 where ŷ i is the estimated value. Torr & Murray repeatedly delete the single point with maximal influence (the point that causes the greatest deviation in the placement of the regression line) thus allowing the fitted model to stay away from the outliers. They refit the regression to the remaining data until there are no more outliers, i.e., the next point with maximal influence lies below a threshold value. Least Squares regression is not robust as the outliers affect the placement of the regression line so it is best suited to outlier diagnostics where the outliers are removed from the next iteration. Figure 6 shows a Least Squares regression line (dashed line) fitted to a data distribution with the outliers A and B present and then again after point A and B have been removed. Although there are only two outliers, they have a considerable affect on the line placement. Torr and Murray (Torr and Murray, 1993) extend the technique for image segmentation. They use a computationally cheap non-robust least median of squares (LMS) regression if the number of outliers is small which minimises equation 5 17 Median n i=1(y i ŷ i ) 2 (5) or a computationally expensive robust random sampling algorithm if the number of outliers is high. LMS is able to accommodate more outliers than LS as it uses the median values. However, random sampling can accommodate larger numbers of outliers which eventually distort LMS. LMS has also been improved to produce the Least Trimmed Squares approach (Rousseeuw and Leroy, 1996) which has faster convergence and minimises equation 6 h ((y i ŷ i ) 2 ) i:n (6) i=1 where (y i ŷ i ) 2 ) 1:n... (y i ŷ i ) 2 ) n:n are the ordered square residuals. The summation function accommodates outliers in the distribution by fitting the regression to only the majority of the data rather than all of the data as in LMS. This region thus depicts normality and LTS highlights outliers as the points with large deviations from the majority. Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.17

18 18 MVE and Convex Peeling aim to compact the convex hull and circumscribe the data with a decision boundary but are only applicable for low dimensional data. Principal Component Analysis (PCA) (Faloutsos et al., 1997)(Parra et al., 1996) in contrast, is suitable for higher dimensional data. It identifies correlated attributes in the data distribution and projects the data onto this lower dimensional subspace. PCA is an unsupervised classifier but is linear and incapable of outperforming the complex non-linear class boundaries identified by the Support Vector Machine (see section 2.4) or neural methods described in section 3. PCA assumes that the subspaces determined by the principal components are compact and this limits its applicability particularly for sparse distributions. However, it is an ideal pre-processor to select a subset of attributes for methods which suffer the Curse of Dimensionality such as the Multi-Layer Perceptron in section 3, proximity-based techniques or symplectic transformations described next. PCA identifies the principal component of greatest variance as each component has an associated eigenvalue whose magnitude corresponds to the variance of the points from the component vector. PCA retains the k principal components with greatest variance and discards all others to preserve maximum information and retain minimal redundancy. Faloutsos et al.(faloutsos et al., 1997) recommend retaining sufficient components so the sum of the eigenvalues of all retained components is at least 85% of the sum of all eigenvalues. They use the principal components to predict attribute values in records by finding the intersection between the given values for the record (i.e., excluding the omitted attribute) and the principal components. If the actual value for an attribute and the predicted value differ then the record is flagged as an outlier. Parra et al. (Parra et al., 1996) have developed a type-3 motor fault detector system which applies PCA to the data and then applies a symplectic transformation to the first few principal components. The symplectic transformation may be used with non-linear data distributions. It maps the input data onto a Gaussian distribution, conserving the volume and separating the training data (normal data) from the outliers. This double transformation preserves information while removing redundancy, generates a density estimation of the data set and thus allows a circular contour of the density to act as the decision boundary. Baker et al. (Baker et al., 1999) employ one of the hierarchical approaches detailed in this survey. The other hierarchical approaches are the decision tree and cluster trees detailed in the machine learning Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.18

19 section. Baker uses a parametric model-based approach for novelty detection in a news story monitor. A hierarchy allows the domain knowledge to be represented at various levels of abstraction so points can be compared for novelty at a fine-grained or less specific level. The hierarchical statistical algorithm induces a topic hierarchy from the word distributions of news stories using Expectation Maximisation (EM) to estimate the parameter settings followed by Deterministic Annealing (DA). DA constructs the hierarchy via maximum likelihood and information theory using a divisive clustering approach to split nodes into sub-nodes, starting from a single cluster, and build the hierarchy top-down. DA stochastically determines the node to split. The system detects novelty when new nodes are added to the hierarchy that represent documents that do not belong to any of the existing event clusters so new events are effectively described by their position in the hierarchy. When EM is used in conjunction with DA it avoids some of the initialisation dependence of EM but at the cost of computational efficiency. DA can avoid local minima which EM is susceptible to but it may produce sub-optimal results Non-Parametric Methods Many Statistical methods described in this section have data-specific parameters ranging from the k values of k-nn and k-means to distance thresholds for the proximity-based approaches to complex model parameters. Other techniques such as those based around convex hulls and regression and the PCA approaches assume the data follows a specific model. These all require aprioridata knowledge. Such information is often not available or is expensive to compute. Many data sets simply do not follow one specific distribution model and are often randomly distributed. Hence, these approaches may be applicable for an outlier detector where all data is accumulated beforehand and may be pre-processed to determine parameter settings or for data where the distribution model is known. Non-parametric approaches, in contrast are more flexible and autonomous. Dasgupta & Forrest (Dasgupta and Forrest, 1996) introduce a nonparametric approach for novelty detection in machinery operation. The authors recognise novelty which contrasts to the other type 3 approaches we describe such as k-means (Nairac et al., 1999) or the ART neural approach (Caudell and Newman, 1993) in section 3 which recognise or classify the normal data space. The machinery operation produces a time-series of real-valued machinery measurements which Dasgupta & Forrest map onto binary vectors using quantisation (binning). The 19 Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.19

20 20 binary vector (string) effectively represents an encoding of the last n real-values from the time series. As the machinery is constantly monitored, new strings (binary vector windows) are generated to represent the current operating characteristics. Dasgupta & Forrest use a set of detectors where all detectors fail to match any strings defining normality (where two strings match if they are identical in a fixed number of contiguous positions (r)). If any detectors match a new string (a new time window of operating characteristics) then a novelty has been detected. The value of r affects the performance of the algorithm and the value must be selected carefully by empirical testing which inevitably slows processing. Each individual recogniser represents a subsection of the input distribution and compares the input. Dasgupta & Forrest s approach would effectively model figure 5 by failing to match any point within the normal boundary but would match any point from outside Semi-Parametric Methods Semi-parametric methods apply local kernel models rather than a single global distribution model. They aim to combine the speed and complexity growth advantage of parametric methods with the model flexibility of non-parametric methods. Kernel-based methods estimate the density distribution of the input space and identify outliers as lying in regions of low density. Tarassenko & Roberts (Roberts and Tarassenko, 1995) and Bishop (Bishop, 1994) use Gaussian Mixture Models to learn a model of normal data by incrementally learning new exemplars. The GMM is represented by equation 7 M p(t x) = α j (x)φ j (t x) (7) j=1 where M is the number of kernels (φ), α j (x) the mixing coefficients, x the input vector and t the target vector. Tarassenko & Roberts classify EEG signatures to detect abnormal signals which represent medical conditions such as epilepsy. In both approaches, each mixture represents a kernel whose width is autonomously determined by the spread of the data. In Bishop s approach the number of mixture models is determined using cross-validation. Tarassenko & Roberts technique adds new mixture models incrementally. If the mixture that best represents the new exemplar is above a threshold distance, then the algorithm adds a new mixture. This distance threshold is determined autonomously during system training. Once training is completed, the final distance threshold represents the novelty threshold for new items to compare against. A Gaussian probability density function is defined Hodge+Austin_OutlierDetection_AIRE381.tex; 19/01/2004; 13:18; p.20

Introduction to partitioning-based clustering methods with a robust example

Introduction to partitioning-based clustering methods with a robust example Reports of the Department of Mathematical Information Technology Series C. Software and Computational Engineering No. C. 1/2006 Introduction to partitioning-based clustering methods with a robust example

More information

Top 10 algorithms in data mining

Top 10 algorithms in data mining Knowl Inf Syst (2008) 14:1 37 DOI 10.1007/s10115-007-0114-2 SURVEY PAPER Top 10 algorithms in data mining Xindong Wu Vipin Kumar J. Ross Quinlan Joydeep Ghosh Qiang Yang Hiroshi Motoda Geoffrey J. McLachlan

More information

2 Basic Concepts and Techniques of Cluster Analysis

2 Basic Concepts and Techniques of Cluster Analysis The Challenges of Clustering High Dimensional Data * Michael Steinbach, Levent Ertöz, and Vipin Kumar Abstract Cluster analysis divides data into groups (clusters) for the purposes of summarization or

More information

Support Vector Data Description

Support Vector Data Description Machine Learning, 54, 45 66, 2004 c 2004 Kluwer Academic Publishers. Manufactured in The Netherlands. Support Vector Data Description DAVID M.J. TAX davidt@first.fhg.de ROBERT P.W. DUIN r.p.w.duin@tnw.tudelft.nl

More information

Introduction to Data Mining and Knowledge Discovery

Introduction to Data Mining and Knowledge Discovery Introduction to Data Mining and Knowledge Discovery Third Edition by Two Crows Corporation RELATED READINGS Data Mining 99: Technology Report, Two Crows Corporation, 1999 M. Berry and G. Linoff, Data Mining

More information

Learning Deep Architectures for AI. Contents

Learning Deep Architectures for AI. Contents Foundations and Trends R in Machine Learning Vol. 2, No. 1 (2009) 1 127 c 2009 Y. Bengio DOI: 10.1561/2200000006 Learning Deep Architectures for AI By Yoshua Bengio Contents 1 Introduction 2 1.1 How do

More information

Representation Learning: A Review and New Perspectives

Representation Learning: A Review and New Perspectives 1 Representation Learning: A Review and New Perspectives Yoshua Bengio, Aaron Courville, and Pascal Vincent Department of computer science and operations research, U. Montreal also, Canadian Institute

More information

What s Strange About Recent Events (WSARE): An Algorithm for the Early Detection of Disease Outbreaks

What s Strange About Recent Events (WSARE): An Algorithm for the Early Detection of Disease Outbreaks Journal of Machine Learning Research 6 (2005) 1961 1998 Submitted 8/04; Revised 3/05; Published 12/05 What s Strange About Recent Events (WSARE): An Algorithm for the Early Detection of Disease Outbreaks

More information

An Introduction to Variable and Feature Selection

An Introduction to Variable and Feature Selection Journal of Machine Learning Research 3 (23) 1157-1182 Submitted 11/2; Published 3/3 An Introduction to Variable and Feature Selection Isabelle Guyon Clopinet 955 Creston Road Berkeley, CA 9478-151, USA

More information

Learning to Select Features using their Properties

Learning to Select Features using their Properties Journal of Machine Learning Research 9 (2008) 2349-2376 Submitted 8/06; Revised 1/08; Published 10/08 Learning to Select Features using their Properties Eyal Krupka Amir Navot Naftali Tishby School of

More information

Data Clustering: A Review

Data Clustering: A Review Data Clustering: A Review AK JAIN Michigan State University MN MURTY Indian Institute of Science AND PJ FLYNN The Ohio State University Clustering is the unsupervised classification of patterns (observations,

More information

THE development of methods for automatic detection

THE development of methods for automatic detection Learning to Detect Objects in Images via a Sparse, Part-Based Representation Shivani Agarwal, Aatif Awan and Dan Roth, Member, IEEE Computer Society 1 Abstract We study the problem of detecting objects

More information

THE PROBLEM OF finding localized energy solutions

THE PROBLEM OF finding localized energy solutions 600 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 45, NO. 3, MARCH 1997 Sparse Signal Reconstruction from Limited Data Using FOCUSS: A Re-weighted Minimum Norm Algorithm Irina F. Gorodnitsky, Member, IEEE,

More information

Data Quality Assessment: A Reviewer s Guide EPA QA/G-9R

Data Quality Assessment: A Reviewer s Guide EPA QA/G-9R United States Office of Environmental EPA/240/B-06/002 Environmental Protection Information Agency Washington, DC 20460 Data Quality Assessment: A Reviewer s Guide EPA QA/G-9R FOREWORD This document is

More information

Distinctive Image Features from Scale-Invariant Keypoints

Distinctive Image Features from Scale-Invariant Keypoints Distinctive Image Features from Scale-Invariant Keypoints David G. Lowe Computer Science Department University of British Columbia Vancouver, B.C., Canada lowe@cs.ubc.ca January 5, 2004 Abstract This paper

More information

Robust Object Detection with Interleaved Categorization and Segmentation

Robust Object Detection with Interleaved Categorization and Segmentation Submission to the IJCV Special Issue on Learning for Vision and Vision for Learning, Sept. 2005, 2 nd revised version Aug. 2007. Robust Object Detection with Interleaved Categorization and Segmentation

More information

Intrusion Detection Techniques for Mobile Wireless Networks

Intrusion Detection Techniques for Mobile Wireless Networks Mobile Networks and Applications? (2003) 1 16 1 Intrusion Detection Techniques for Mobile Wireless Networks Yongguang Zhang HRL Laboratories LLC, Malibu, California E-mail: ygz@hrl.com Wenke Lee College

More information



More information

Cluster Ensembles A Knowledge Reuse Framework for Combining Multiple Partitions

Cluster Ensembles A Knowledge Reuse Framework for Combining Multiple Partitions Journal of Machine Learning Research 3 (2002) 583-67 Submitted 4/02; Published 2/02 Cluster Ensembles A Knowledge Reuse Framework for Combining Multiple Partitions Alexander Strehl alexander@strehl.com

More information

Generating Human-like Motion for Robots

Generating Human-like Motion for Robots Generating Human-like Motion for Robots Michael J. Gielniak, C. Karen Liu, and Andrea L. Thomaz Abstract Action prediction and fluidity are a key elements of human-robot teamwork. If a robot s actions

More information

High-Dimensional Image Warping

High-Dimensional Image Warping Chapter 4 High-Dimensional Image Warping John Ashburner & Karl J. Friston The Wellcome Dept. of Imaging Neuroscience, 12 Queen Square, London WC1N 3BG, UK. Contents 4.1 Introduction.................................

More information

Why Does Unsupervised Pre-training Help Deep Learning?

Why Does Unsupervised Pre-training Help Deep Learning? Journal of Machine Learning Research 11 (2010) 625-660 Submitted 8/09; Published 2/10 Why Does Unsupervised Pre-training Help Deep Learning? Dumitru Erhan Yoshua Bengio Aaron Courville Pierre-Antoine Manzagol

More information

Gene Selection for Cancer Classification using Support Vector Machines

Gene Selection for Cancer Classification using Support Vector Machines Gene Selection for Cancer Classification using Support Vector Machines Isabelle Guyon+, Jason Weston+, Stephen Barnhill, M.D.+ and Vladimir Vapnik* +Barnhill Bioinformatics, Savannah, Georgia, USA * AT&T

More information



More information

Indexing by Latent Semantic Analysis. Scott Deerwester Graduate Library School University of Chicago Chicago, IL 60637

Indexing by Latent Semantic Analysis. Scott Deerwester Graduate Library School University of Chicago Chicago, IL 60637 Indexing by Latent Semantic Analysis Scott Deerwester Graduate Library School University of Chicago Chicago, IL 60637 Susan T. Dumais George W. Furnas Thomas K. Landauer Bell Communications Research 435

More information


EVALUATION OF GAUSSIAN PROCESSES AND OTHER METHODS FOR NON-LINEAR REGRESSION. Carl Edward Rasmussen EVALUATION OF GAUSSIAN PROCESSES AND OTHER METHODS FOR NON-LINEAR REGRESSION Carl Edward Rasmussen A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy, Graduate

More information

Large Scale Online Learning of Image Similarity Through Ranking

Large Scale Online Learning of Image Similarity Through Ranking Journal of Machine Learning Research 11 (21) 119-1135 Submitted 2/9; Revised 9/9; Published 3/1 Large Scale Online Learning of Image Similarity Through Ranking Gal Chechik Google 16 Amphitheatre Parkway

More information



More information

Foundations of Data Science 1

Foundations of Data Science 1 Foundations of Data Science John Hopcroft Ravindran Kannan Version 2/8/204 These notes are a first draft of a book being written by Hopcroft and Kannan and in many places are incomplete. However, the notes

More information

THE adoption of classical statistical modeling techniques

THE adoption of classical statistical modeling techniques 236 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 28, NO. 2, FEBRUARY 2006 Data Driven Image Models through Continuous Joint Alignment Erik G. Learned-Miller Abstract This paper

More information